oota-llvm.git
11 years ago[SystemZ] Add DebugInfo test cases
Ulrich Weigand [Mon, 6 May 2013 16:18:29 +0000 (16:18 +0000)]
[SystemZ] Add DebugInfo test cases

This adds all DebugInfo tests for the SystemZ target.

This version of the patch incorporates feedback from reviews by
Eric Christopher and Rafael Espindola.  Thanks to all reviewers!

Patch by Richard Sandiford.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181205 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add CodeGen test cases
Ulrich Weigand [Mon, 6 May 2013 16:17:29 +0000 (16:17 +0000)]
[SystemZ] Add CodeGen test cases

This adds all CodeGen tests for the SystemZ target.

This version of the patch incorporates feedback from a review by
Sean Silva.  Thanks to all reviewers!

Patch by Richard Sandiford.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181204 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add back end
Ulrich Weigand [Mon, 6 May 2013 16:15:19 +0000 (16:15 +0000)]
[SystemZ] Add back end

This adds the actual lib/Target/SystemZ target files necessary to
implement the SystemZ target.  Note that at this point, the target
cannot yet be built since the configure bits are missing.  Those
will be provided shortly by a follow-on patch.

This version of the patch incorporates feedback from reviews by
Chris Lattner and Anton Korobeynikov.  Thanks to all reviewers!

Patch by Richard Sandiford.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181203 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Define DWARF encoding
Ulrich Weigand [Mon, 6 May 2013 16:11:12 +0000 (16:11 +0000)]
[SystemZ] Define DWARF encoding

This is another patch in preparation for adding the SystemZ target.
It defines the appropriate values for DWARF encodings; the intent
is to be compatible with what GCC currently does on the target.

Patch by Richard Sandiford.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181201 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSimplify JIT unit test #ifdefs
Ulrich Weigand [Mon, 6 May 2013 16:10:35 +0000 (16:10 +0000)]
Simplify JIT unit test #ifdefs

Several platforms need to disable all old-JIT unit tests, since they only
support the new MCJIT.  This currently done via #ifdef'ing out those tests
in the ExecutionEngine/JIT/*.cpp files.  As those #ifdef's have grown
historically, we now have a number of repeated directives which -in total-
cover nearly the whole file, but leave a couple of helper functions out.
When building the tests with clang itself, those helper functions now
cause spurious "unused function" warnings.

To fix those warnings, and also to remove the duplicate #ifdef conditions
and make it easier to disable the tests for a new target, this patch
consolidates the #ifdefs into a single one per file, which covers all
the tests including all helper routines.

Tested on PowerPC and SystemZ.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181200 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFree the exception object. Should fix the vg bots.
Rafael Espindola [Mon, 6 May 2013 13:30:52 +0000 (13:30 +0000)]
Free the exception object. Should fix the vg bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181195 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Fix memory corruption in AsmParser
Ulrich Weigand [Mon, 6 May 2013 11:16:57 +0000 (11:16 +0000)]
[PowerPC] Fix memory corruption in AsmParser

As pointed out by Evgeniy Stepanov, assigning a std::string temporary
to a StringRef is not a good idea.  Rework MatchRegisterName to avoid
using the .lower routine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181192 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix formatting. Patch by o11c.
Duncan Sands [Mon, 6 May 2013 08:55:45 +0000 (08:55 +0000)]
Fix formatting.  Patch by o11c.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181189 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix slightly too aggressive conact_vector optimization.
Michael Kuperstein [Mon, 6 May 2013 08:06:13 +0000 (08:06 +0000)]
Fix slightly too aggressive conact_vector optimization.
(Would sometimes optimize away conacts used to extend a vector with undef values)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181186 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd a testcase that checks that we generate functions with frame
Bill Wendling [Mon, 6 May 2013 05:45:57 +0000 (05:45 +0000)]
Add a testcase that checks that we generate functions with frame
pointers or not depending upon the function attributes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181180 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoXFAIL for cygwin.
Rafael Espindola [Mon, 6 May 2013 03:35:56 +0000 (03:35 +0000)]
XFAIL for cygwin.

Looks like symbol resolution is not working on cygwin, the test fails
because __gxx_personality_v0 is not found.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181179 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUpdate the comment to mention that we use TTI.
Nadav Rotem [Mon, 6 May 2013 03:06:36 +0000 (03:06 +0000)]
Update the comment to mention that we use TTI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181178 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert r164763 because it introduces new shuffles.
Nadav Rotem [Mon, 6 May 2013 02:39:09 +0000 (02:39 +0000)]
Revert r164763 because it introduces new shuffles.
Thanks Nick Lewycky for pointing this out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181177 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix unchecked uses of DominatorTree in MemoryDependenceAnalysis.
Matt Arsenault [Mon, 6 May 2013 02:07:24 +0000 (02:07 +0000)]
Fix unchecked uses of DominatorTree in MemoryDependenceAnalysis.

Use unknown results for places where it would be needed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181176 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix const merging when an alias of a const is llvm.used.
Rafael Espindola [Mon, 6 May 2013 01:48:55 +0000 (01:48 +0000)]
Fix const merging when an alias of a const is llvm.used.

We used to disable constant merging not only if a constant is llvm.used, but
also if an alias of a constant is llvm.used. This change fixes that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181175 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoThis should also fail on ARM.
Rafael Espindola [Sun, 5 May 2013 22:42:34 +0000 (22:42 +0000)]
This should also fail on ARM.

We currently have no way to register new eh frames on ARM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181172 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix XFAIL line.
Rafael Espindola [Sun, 5 May 2013 21:30:10 +0000 (21:30 +0000)]
Fix XFAIL line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181171 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoXFAIL this on ppc64.
Rafael Espindola [Sun, 5 May 2013 21:04:18 +0000 (21:04 +0000)]
XFAIL this on ppc64.

It looks like eh uses an unimplemented relocation on pp64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181169 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPort ExceptionDemo to MCJIT.
Rafael Espindola [Sun, 5 May 2013 20:57:58 +0000 (20:57 +0000)]
Port ExceptionDemo to MCJIT.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181168 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd EH support to the MCJIT.
Rafael Espindola [Sun, 5 May 2013 20:43:10 +0000 (20:43 +0000)]
Add EH support to the MCJIT.

This gets exception handling working on ELF and Macho (x86-64 at least).
Other than the EH frame registration, this patch also implements support
for GOT relocations which are used to locate the personality function on
MachO.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181167 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTest case for r181160 and r181161. rdar://13782395
Evan Cheng [Sun, 5 May 2013 18:07:15 +0000 (18:07 +0000)]
Test case for r181160 and r181161. rdar://13782395

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181162 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM AnalyzeBranch should conservatively return true when it sees a predicated
Evan Cheng [Sun, 5 May 2013 18:06:32 +0000 (18:06 +0000)]
ARM AnalyzeBranch should conservatively return true when it sees a predicated
indirect branch at the end of the BB. Otherwise if-converter, branch folding
pass may incorrectly update its successor info if it consider BB as fallthrough
to the next BB.

rdar://13782395

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181161 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTeach if-converter to avoid removing BBs whose addresses are takne. rdar://13782395
Evan Cheng [Sun, 5 May 2013 18:03:49 +0000 (18:03 +0000)]
Teach if-converter to avoid removing BBs whose addresses are takne. rdar://13782395

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181160 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLoopVectorize: Print values instead of pointers in debug output.
Benjamin Kramer [Sun, 5 May 2013 14:54:52 +0000 (14:54 +0000)]
LoopVectorize: Print values instead of pointers in debug output.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181157 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[docs] Update Target Feature Matrix for the XCore backend.
Richard Osborne [Sun, 5 May 2013 14:09:55 +0000 (14:09 +0000)]
[docs] Update Target Feature Matrix for the XCore backend.

Disassembler support has recently been added. Fill in some other unknowns
at the same time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181156 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Add LDAPB instructions.
Richard Osborne [Sun, 5 May 2013 13:36:53 +0000 (13:36 +0000)]
[XCore] Add LDAPB instructions.

With the change the disassembler now supports the XCore ISA in its
entirety.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181155 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Update LDAP to use pcrel_imm.
Richard Osborne [Sun, 5 May 2013 13:33:10 +0000 (13:33 +0000)]
[XCore] Update LDAP to use pcrel_imm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181154 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Rename calltarget -> pcrel_imm.
Richard Osborne [Sun, 5 May 2013 13:29:02 +0000 (13:29 +0000)]
[XCore] Rename calltarget -> pcrel_imm.

No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181153 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Add BLRB instructions.
Richard Osborne [Sun, 5 May 2013 13:24:16 +0000 (13:24 +0000)]
[XCore] Add BLRB instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181152 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Remove '-' from back branch asm syntax.
Richard Osborne [Sun, 5 May 2013 13:20:22 +0000 (13:20 +0000)]
[XCore] Remove '-' from back branch asm syntax.

Instead operands are treated as negative immediates
where the sign bit is implicit in the instruction
encoding.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181151 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoInlineSpiller: Remove quadratic behavior.
Benjamin Kramer [Sun, 5 May 2013 11:29:14 +0000 (11:29 +0000)]
InlineSpiller: Remove quadratic behavior.

No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181149 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFor ARM backend, fixed "byval" attribute support.
Stepan Dyatkovskiy [Sun, 5 May 2013 07:48:36 +0000 (07:48 +0000)]
For ARM backend, fixed "byval" attribute support.
Now even the small structures could be passed within byval (small enough
to be stored in GPRs).
In regression tests next function prototypes are checked:

PR15293:
  %artz = type { i32 }
  define void @foo(%artz* byval %s)
  define void @foo2(%artz* byval %s, i32 %p, %artz* byval %s2)
foo: "s" stored in R0
foo2: "s" stored in R0, "s2" stored in R2.

Next AAPCS rules are checked:
5.5 Parameters Passing, C.4 and C.5,
"ParamSize" is parameter size in 32bit words:
-- NSAA != 0, NCRN < R4 and NCRN+ParamSize > R4.
   Parameter should be sent to the stack; NCRN := R4.
-- NSAA != 0, and NCRN < R4, NCRN+ParamSize < R4.
   Parameter stored in GPRs; NCRN += ParamSize.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181148 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd missing PatternMatch.cpp to CMakeLists.txt
Arnold Schwaighofer [Sun, 5 May 2013 02:14:28 +0000 (02:14 +0000)]
Add missing PatternMatch.cpp to CMakeLists.txt

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181147 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPatternMatch: Fix documentation - 'function' not 'attribute'
Arnold Schwaighofer [Sun, 5 May 2013 02:01:41 +0000 (02:01 +0000)]
PatternMatch: Fix documentation - 'function' not 'attribute'

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181146 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove a recently redundant transform from X86ISelLowering.
David Majnemer [Sun, 5 May 2013 02:00:10 +0000 (02:00 +0000)]
Remove a recently redundant transform from X86ISelLowering.

X86ISelLowering has support to treat:
(icmp ne (and (xor %flags, -1), (shl 1, flag)), 0)

as if it were actually:
(icmp eq (and %flags, (shl 1, flag)), 0)

However, r179386 has code at the InstCombine level to handle this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181145 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLoopVectorize: Add support for floating point min/max reductions
Arnold Schwaighofer [Sun, 5 May 2013 01:54:48 +0000 (01:54 +0000)]
LoopVectorize: Add support for floating point min/max reductions

Add support for min/max reductions when "no-nans-float-math" is enabled. This
allows us to assume we have ordered floating point math and treat ordered and
unordered predicates equally.

radar://13723044

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181144 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPatternMatch: Matcher for (un)ordered floating point min/max
Arnold Schwaighofer [Sun, 5 May 2013 01:54:46 +0000 (01:54 +0000)]
PatternMatch: Matcher for (un)ordered floating point min/max

Add support for matching 'ordered' and 'unordered' floating point min/max
constructs.

In LLVM we can express min/max functions as a combination of compare and select.
We have support for matching such constructs for integers but not for floating
point. In floating point math there is no total order because of the presence of
'NaN'. Therefore, we have to be careful to preserve the original fcmp semantics
when interpreting floating point compare select combinations as a minimum or
maximum function. The resulting 'ordered/unordered' floating point maximum
function has to select the same value as the select/fcmp combination it is based
on.

 ordered_max(x,y)   = max(x,y) iff x and y are not NaN, y otherwise
 unordered_max(x,y) = max(x,y) iff x and y are not NaN, x otherwise
 ordered_min(x,y)   = min(x,y) iff x and y are not NaN, y otherwise
 unordered_min(x,y) = min(x,y) iff x and y are not NaN, x otherwise

This matches the behavior of the underlying select(fcmp(olt/ult/.., L, R), L, R)
construct.

Any code using this predicate has to preserve this semantics.

A follow-up patch will use this to implement floating point min/max reductions
in the vectorizer.

radar://13723044

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181143 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLoopVectorizer: Cleanup of miminimum/maximum pattern match code
Arnold Schwaighofer [Sun, 5 May 2013 01:54:44 +0000 (01:54 +0000)]
LoopVectorizer: Cleanup of miminimum/maximum pattern match code

No need for setting the operands. The pointers are going to be bound by the
matcher.

radar://13723044

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181142 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLoopVectorize: We don't need an identity element for min/max reductions
Arnold Schwaighofer [Sun, 5 May 2013 01:54:42 +0000 (01:54 +0000)]
LoopVectorize: We don't need an identity element for min/max reductions

We can just use the initial element that feeds the reduction.

  max(max(x, y), z) == max(max(x,y), max(x,z))

radar://13723044

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181141 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd ArrayRef constructor from None, and do the cleanups that this constructor enables
Dmitri Gribenko [Sun, 5 May 2013 00:40:33 +0000 (00:40 +0000)]
Add ArrayRef constructor from None, and do the cleanups that this constructor enables

Patch by Robert Wilhelm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181138 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agowhitespace
Nadav Rotem [Sat, 4 May 2013 23:27:32 +0000 (23:27 +0000)]
whitespace

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181137 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix an odd comment.
Nadav Rotem [Sat, 4 May 2013 23:24:56 +0000 (23:24 +0000)]
Fix an odd comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181136 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: enable MCJIT and tests now that everything passes.
Tim Northover [Sat, 4 May 2013 20:14:22 +0000 (20:14 +0000)]
AArch64: enable MCJIT and tests now that everything passes.

This removes dire warnings about AArch64 being unsupported and enables
the tests when appropriate on this platform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181135 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: implement 64-bit absolute relocation in MCJIT
Tim Northover [Sat, 4 May 2013 20:14:14 +0000 (20:14 +0000)]
AArch64: implement 64-bit absolute relocation in MCJIT

This is about the simplest relocation, but surprisingly rare in actual
code.

It occurs in (for example) the MCJIT test test-ptr-reloc.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181134 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: add stubs to support long function calls on MCJIT
Tim Northover [Sat, 4 May 2013 20:14:09 +0000 (20:14 +0000)]
AArch64: add stubs to support long function calls on MCJIT

As with global accesses, external functions could exist anywhere in
memory. Therefore the stub must create a complete 64-bit address. This
patch implements the fragment as (roughly):
    movz x16, #:abs_g3:somefunc
    movk x16, #:abs_g2_nc:somefunc
    movk x16, #:abs_g1_nc:somefunc
    movk x16, #:abs_g0_nc:somefunc
    br x16

In principle we could save 4 bytes by using a literal-load instead,
but it is unclear that would be more efficient and can only be tested
when real hardware is readily available.

This allows (for example) the MCJIT test 2003-05-07-ArgumentTest to
pass on AArch64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181133 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: implement relocations for global access
Tim Northover [Sat, 4 May 2013 20:14:04 +0000 (20:14 +0000)]
AArch64: implement relocations for global access

The large memory model (default and main viable for JIT) emits
addresses in need of relocation as
    movz x0, #:abs_g3:somewhere
    movk x0, #:abs_g2_nc:somewhere
    movk x0, #:abs_g1_nc:somewhere
    movk x0, #:abs_g0_nc:somewhere

To support this we must implement those four relocations in the
dynamic loader.

This allows (for example) the test-global.ll MCJIT test to pass on
AArch64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181132 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: implement first relocation required for MCJIT
Tim Northover [Sat, 4 May 2013 20:13:59 +0000 (20:13 +0000)]
AArch64: implement first relocation required for MCJIT

R_AARCH64_PCREL32 is present in even trivial .eh_frame sections and so
is required to compile any function without the "nounwind" attribute.

This change implements very basic infrastructure in the RuntimeDyldELF
file and allows (for example) the test-shift.ll MCJIT test to pass
on AArch64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181131 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoBuild system changes to enable MCJIT on AArch64
Tim Northover [Sat, 4 May 2013 20:13:52 +0000 (20:13 +0000)]
Build system changes to enable MCJIT on AArch64

These changes just allow AArch64 to take part in the MCJIT world when
built correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181130 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: use __clear_cache under GCCish environments
Tim Northover [Sat, 4 May 2013 18:52:44 +0000 (18:52 +0000)]
AArch64: use __clear_cache under GCCish environments

AArch64 is going to need some kind of cache-invalidation in order to
successfully JIT since it has a weak memory-model. This is provided by
a __clear_cache builtin in libgcc, which acts very much like the
32-bit ARM equivalent (on platforms where it exists).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181129 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix buildbot failure on 64 bit linux due to std::max() having different
Richard Osborne [Sat, 4 May 2013 17:41:01 +0000 (17:41 +0000)]
Fix buildbot failure on 64 bit linux due to std::max() having different
operand types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181128 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Remove unused operand type.
Richard Osborne [Sat, 4 May 2013 17:30:05 +0000 (17:30 +0000)]
[XCore] Remove unused operand type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181127 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Make use of the target independent global address offset folding.
Richard Osborne [Sat, 4 May 2013 17:24:33 +0000 (17:24 +0000)]
[XCore] Make use of the target independent global address offset folding.

This let us to remove some custom code that matched constant offsets
from globals at instruction selection time as a special addressing mode.
No intended functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181126 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Simplify code that checks for an aligned base plus a constant.
Richard Osborne [Sat, 4 May 2013 17:17:10 +0000 (17:17 +0000)]
[XCore] Simplify code that checks for an aligned base plus a constant.

The code now makes use of ComputeMaskedBits,
SelectionDAG::isBaseWithConstantOffset and TargetLowering::isGAPlusOffset
where appropriate reducing the amount of logic needed in XCoreISelLowering.
No intended functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181125 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Move lowering of thread local storage to a separate pass.
Richard Osborne [Sat, 4 May 2013 17:01:55 +0000 (17:01 +0000)]
[XCore] Move lowering of thread local storage to a separate pass.

Thread local storage is not supported by the XMOS linker so we handle
thread local variables by lowering the variable to an array of n elements
(where n is the number of hardware threads per core, currently 8
for all XMOS devices) indexed by the the current thread ID.

Previously this lowering was spread across the XCoreISelLowering and the
XCoreAsmPrinter classes. Moving this to a separate pass should be much
cleaner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181124 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: assert code model is small for TLS accesses
Tim Northover [Sat, 4 May 2013 16:54:11 +0000 (16:54 +0000)]
AArch64: assert code model is small for TLS accesses

Supporting TLS in the large memory model is rather difficult at the
moment, so make sure no-one gets into difficulties by mistake.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181121 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: support literal pool access in large memory model.
Tim Northover [Sat, 4 May 2013 16:54:07 +0000 (16:54 +0000)]
AArch64: support literal pool access in large memory model.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181120 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: support large code model for jump-tables
Tim Northover [Sat, 4 May 2013 16:54:00 +0000 (16:54 +0000)]
AArch64: support large code model for jump-tables

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181119 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: implement support for blockaddress in large code model
Tim Northover [Sat, 4 May 2013 16:53:53 +0000 (16:53 +0000)]
AArch64: implement support for blockaddress in large code model

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181118 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: implement large code model access to global variables.
Tim Northover [Sat, 4 May 2013 16:53:46 +0000 (16:53 +0000)]
AArch64: implement large code model access to global variables.

The MOVZ/MOVK instruction sequence may not be the most efficient (a
literal-pool load could be better) but adding that would require
reinstating the ConstantIslands pass.

For now the sequence is correct, and that's enough. Beware, as of
commit GNU ld does not appear to support the relocations needed for
this. Its primary purpose (for now) will be to support JITed code,
since in that case there is no guarantee of where your code will end
up in memory relative to external symbols it references.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181117 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[XCore] Use static relocation model by default.
Richard Osborne [Sat, 4 May 2013 16:40:58 +0000 (16:40 +0000)]
[XCore] Use static relocation model by default.

This allows us to get get rid of a hack in XCoreTargetObjectFile where the
the DataRel* sections were overridden.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181116 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAllow host triple to be correctly overridden in CMake builds
Tim Northover [Sat, 4 May 2013 07:36:23 +0000 (07:36 +0000)]
Allow host triple to be correctly overridden in CMake builds

The intended semantics mirror autoconf, where the user is able to
specify a host triple, but if it's left to the build system then
"config.guess" is invoked for the default.

This also renames the LLVM_HOSTTRIPLE define to LLVM_HOST_TRIPLE to
fit in with the style of the surrounding defines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181112 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoOptimize llvm-link too.
Rafael Espindola [Sat, 4 May 2013 05:30:49 +0000 (05:30 +0000)]
Optimize llvm-link too.

This takes the linking of almost all modules in a clang build from 6:32
to 0:19.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181105 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a performance bug in the Linker.
Rafael Espindola [Sat, 4 May 2013 05:05:18 +0000 (05:05 +0000)]
Fix a performance bug in the Linker.

Now that we hava a convinient place to keep it, remeber the set of
identified structs as we merge modules.

This speeds up the linking of all the bitcode files in clang with the
gold plugin and -plugin-opt=emit-llvm (i.e., link only, no codegen) from
5:25 minutes to 13.6 seconds!

Patch by Xiaofei Wan!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181104 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoImplement Linker::LinkModules with Linker::linkInModule.
Rafael Espindola [Sat, 4 May 2013 04:08:02 +0000 (04:08 +0000)]
Implement Linker::LinkModules with Linker::linkInModule.

Flipping which one is the implementation will let us optimize linkInModule.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181102 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoNow that Linker.cpp is almost empty, merge it into LinkModules.cpp.
Rafael Espindola [Sat, 4 May 2013 03:48:37 +0000 (03:48 +0000)]
Now that Linker.cpp is almost empty, merge it into LinkModules.cpp.

Also remove unused includes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181100 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLast batch of cleanups to Linker.h.
Rafael Espindola [Sat, 4 May 2013 03:06:50 +0000 (03:06 +0000)]
Last batch of cleanups to Linker.h.

Update comments, fix * placement, fix method names that are not
used in clang, add a linkInModule that takes a Mode and put it
in Linker.cpp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181099 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDon't construct or delete a module on the Linker.
Rafael Espindola [Sat, 4 May 2013 02:43:00 +0000 (02:43 +0000)]
Don't construct or delete a module on the Linker.

The linker is now responsible only for actually linking the modules, it
is up to the clients to create and destroy them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181098 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDon't store the context in the Linker.
Rafael Espindola [Sat, 4 May 2013 02:34:41 +0000 (02:34 +0000)]
Don't store the context in the Linker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181097 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove unused members and constructor arguments.
Rafael Espindola [Sat, 4 May 2013 02:28:57 +0000 (02:28 +0000)]
Remove unused members and constructor arguments.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181096 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd missing header.
Rafael Espindola [Sat, 4 May 2013 02:21:46 +0000 (02:21 +0000)]
Add missing header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181095 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDelete dead code from the linker.
Rafael Espindola [Sat, 4 May 2013 02:13:18 +0000 (02:13 +0000)]
Delete dead code from the linker.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181094 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse consistent function names.
Krzysztof Parzyszek [Sat, 4 May 2013 01:30:49 +0000 (01:30 +0000)]
Use consistent function names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181090 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTabs to spaces. No functionality change.
Nick Lewycky [Sat, 4 May 2013 01:08:15 +0000 (01:08 +0000)]
Tabs to spaces. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181082 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert r181009.
Amara Emerson [Fri, 3 May 2013 23:57:17 +0000 (23:57 +0000)]
Revert r181009.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181079 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove some uneeded pseudos in the presence of the naked function attribute.
Reed Kotler [Fri, 3 May 2013 23:17:24 +0000 (23:17 +0000)]
Remove some uneeded pseudos in the presence of the naked function attribute.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181072 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDelete test instead.
Amara Emerson [Fri, 3 May 2013 22:39:03 +0000 (22:39 +0000)]
Delete test instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181066 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTemporarily disable failing test.
Amara Emerson [Fri, 3 May 2013 22:27:48 +0000 (22:27 +0000)]
Temporarily disable failing test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181062 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove this hack. We can support this better with function attributes.
Bill Wendling [Fri, 3 May 2013 21:53:50 +0000 (21:53 +0000)]
Remove this hack. We can support this better with function attributes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181059 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Avoid using '$' in generated assembler code
Ulrich Weigand [Fri, 3 May 2013 19:53:04 +0000 (19:53 +0000)]
[PowerPC] Avoid using '$' in generated assembler code

PowerPC assemblers are supposed to support a stand-alone '$' symbol
as an alternative of '.' to refer to the current PC.  This does not
work in the LLVM assembler parser yet.

To avoid bootstrap failures when using the LLVM assembler as system
assembler, this patch modifies the assembler source code generated
by LLVM to avoid using '$' (and simply use '.' instead).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181054 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Parse platform-specifc variant kinds in AsmParser
Ulrich Weigand [Fri, 3 May 2013 19:52:35 +0000 (19:52 +0000)]
[PowerPC] Parse platform-specifc variant kinds in AsmParser

This patch adds support for PowerPC platform-specific variant
kinds in MCSymbolRefExpr::getVariantKindForName, and also
adds a test case to verify they are translated to the appropriate
fixup type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181053 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Add some Book II instructions to AsmParser
Ulrich Weigand [Fri, 3 May 2013 19:51:09 +0000 (19:51 +0000)]
[PowerPC] Add some Book II instructions to AsmParser

This patch adds a couple of Book II instructions (isync, icbi) to the
PowerPC assembler parser.  These are needed when bootstrapping clang
with the integrated assembler forced on, because they are used in
inline asm statements in the code base.

The test case adds the full list of Book II storage control instructions,
including associated extended mnemonics.  Again, those that are not yet
supported as marked as FIXME.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181052 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Support extended mnemonics in AsmParser
Ulrich Weigand [Fri, 3 May 2013 19:50:27 +0000 (19:50 +0000)]
[PowerPC] Support extended mnemonics in AsmParser

This patch adds infrastructure to support extended mnemonics in the
PowerPC assembler parser.  It adds support specifically for those
extended mnemonics that LLVM will itself generate.

The test case lists *all* extended mnemonics according to the
PowerPC ISA v2.06 Book I, but marks those not yet supported
as FIXME.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181051 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[PowerPC] Add assembler parser
Ulrich Weigand [Fri, 3 May 2013 19:49:39 +0000 (19:49 +0000)]
[PowerPC] Add assembler parser

This adds assembler parser support to the PowerPC back end.

The parser will run for any powerpc-*-* and powerpc64-*-* triples,
but was tested only on 64-bit Linux.  The supported syntax is
intended to be compatible with the GNU assembler.

The parser does not yet support all PowerPC instructions, but
it does support anything that is generated by LLVM itself.
There is no support for testing restricted instruction sets yet,
i.e. the parser will always accept any instructions it knows,
no matter what feature flags are given.

Instruction operands will be checked for validity and errors
generated.  (Error handling in general could still be improved.)

The patch adds a number of test cases to verify instruction
and operand encodings.  The tests currently cover all instructions
from the following PowerPC ISA v2.06 Book I facilities:
Branch, Fixed-point, Floating-Point, and Vector.
Note that a number of these instructions are not yet supported
by the back end; they are marked with FIXME.

A number of follow-on check-ins will add extra features.  When
they are all included, LLVM passes all tests (including bootstrap)
when using clang -cc1as as the system assembler.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181050 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDecompose GVN::processNonLocalLoad() (about 400 LOC) into smaller helper functions...
Shuxin Yang [Fri, 3 May 2013 19:17:26 +0000 (19:17 +0000)]
Decompose GVN::processNonLocalLoad() (about 400 LOC) into smaller helper functions. No function change.

This function consists of following steps:
   1. Collect dependent memory accesses.
   2. Analyze availability.
   3. Perform fully redundancy elimination, or
   4. Perform PRE, depending on the availability

 Step 2, 3 and 4 are now moved to three helper routines.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181047 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Split the DSP control register and define one register for each field of
Akira Hatanaka [Fri, 3 May 2013 18:37:49 +0000 (18:37 +0000)]
[mips] Split the DSP control register and define one register for each field of
its fields.

This removes false dependencies between DSP instructions which access different
fields of the the control register. Implicit register operands are added to
instructions RDDSP and WRDSP after instruction selection, depending on the
value of the mask operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181041 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLoopVectorizer: Add support for if-conversion of PHINodes with 3+ incoming values.
Nadav Rotem [Fri, 3 May 2013 17:42:55 +0000 (17:42 +0000)]
LoopVectorizer: Add support for if-conversion of PHINodes with 3+ incoming values.
By supporting the vectorization of PHINodes with more than two incoming values we can increase the complexity of nested if statements.

We can now vectorize this loop:

int foo(int *A, int *B, int n) {
  for (int i=0; i < n; i++) {
    int x = 9;
    if (A[i] > B[i]) {
      if (A[i] > 19) {
        x = 3;
      } else if (B[i] < 4 ) {
        x = 4;
      } else {
        x = 5;
      }
    }
    A[i] = x;
  }
}

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181037 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Expand vector or, shl, srl, and xor nodes
Tom Stellard [Fri, 3 May 2013 17:21:31 +0000 (17:21 +0000)]
R600: Expand vector or, shl, srl, and xor nodes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181035 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: BFI_INT is a vector-only instruction
Tom Stellard [Fri, 3 May 2013 17:21:24 +0000 (17:21 +0000)]
R600: BFI_INT is a vector-only instruction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181034 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Add pattern for SHA-256 Ma function
Tom Stellard [Fri, 3 May 2013 17:21:20 +0000 (17:21 +0000)]
R600: Add pattern for SHA-256 Ma function

This can be optimized using the BFI_INT instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181033 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoR600: Clean up comments in Processors.td
Tom Stellard [Fri, 3 May 2013 17:21:14 +0000 (17:21 +0000)]
R600: Clean up comments in Processors.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181032 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAutoconf: Compile cxxabi.h in C++ mode.
Benjamin Kramer [Fri, 3 May 2013 15:55:06 +0000 (15:55 +0000)]
Autoconf: Compile cxxabi.h in C++ mode.

Should fix PR15877.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181026 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRegionInfo: Do not crash if unreachable block is found
Tobias Grosser [Fri, 3 May 2013 15:48:34 +0000 (15:48 +0000)]
RegionInfo: Do not crash if unreachable block is found

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181025 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCMake: cxxabi.h is only available to C++ compilers, use the right check macro.
Benjamin Kramer [Fri, 3 May 2013 14:48:29 +0000 (14:48 +0000)]
CMake: cxxabi.h is only available to C++ compilers, use the right check macro.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181019 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove comment that no target supports 128-bit IEEE floats
Richard Sandiford [Fri, 3 May 2013 14:32:27 +0000 (14:32 +0000)]
Remove comment that no target supports 128-bit IEEE floats

The soon-to-be-committed SystemZ port uses 128-bit IEEE floats.
MIPS64 GNU/Linux does too (albeit with unusual NaNs).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181016 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add MCJIT support
Richard Sandiford [Fri, 3 May 2013 14:15:35 +0000 (14:15 +0000)]
[SystemZ] Add MCJIT support

Another step towards reinstating the SystemZ backend.  I'll commit
the configure changes separately (TARGET_HAS_JIT etc.), then commit
a patch to enable the MCJIT tests on SystemZ.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181015 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Support System Z as host architecture
Ulrich Weigand [Fri, 3 May 2013 12:22:11 +0000 (12:22 +0000)]
[SystemZ] Support System Z as host architecture

The llvm::sys::AddSignalHandler function (as well as related routines) in
lib/Support/Unix/Signals.inc currently registers a signal handler routine
via "sigaction".  When this handler is called due to a SIGSEGV, SIGILL or
similar signal, it will show a stack backtrace, deactivate the handler,
and then simply return to the operating system.  The intent is that the
OS will now retry execution at the same location as before, which ought
to again trigger the same error condition and cause the same signal to be
delivered again.  Since the hander is now deactivated, the OS will take
its default action (usually, terminate the program and possibly create
a core dump).

However, this method doesn't work reliably on System Z:  With certain
signals (namely SIGILL, SIGFPE, and SIGTRAP), the program counter stored
by the kernel on the signal stack frame (which is the location where
execution will resume) is not the instruction that triggered the fault,
but then instruction *after it*.  When the LLVM signal handler simply
returns to the kernel, execution will then resume at *that* address,
which will not trigger the problem again, but simply go on and execute
potentially unrelated code leading to random errors afterwards.

To fix this, the patch simply goes and re-raises the signal in question
directly from the handler instead of returning from it.  This is done
only on System Z and only for those signals that have this particular
problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181010 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for reading ARM ELF build attributes.
Amara Emerson [Fri, 3 May 2013 11:36:35 +0000 (11:36 +0000)]
Add support for reading ARM ELF build attributes.

Build attribute sections can now be read if they exist via ELFObjectFile, and
the llvm-readobj tool has been extended with an option to dump this information
if requested. Regression tests are also included which exercise these features.

Also update the docs with a fixed ARM ABI link and a new link to the Addenda
which provides the build attributes specification.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181009 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add ELF relocation support
Richard Sandiford [Fri, 3 May 2013 11:11:15 +0000 (11:11 +0000)]
[SystemZ] Add ELF relocation support

Another step towards reinstating the SystemZ backend.  Tests will be
included in the main backend patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181008 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[SystemZ] Add llvm::Triple::systemz
Richard Sandiford [Fri, 3 May 2013 11:05:17 +0000 (11:05 +0000)]
[SystemZ] Add llvm::Triple::systemz

First step towards reinstating the SystemZ backend.  Tests will be
included in the main backend patch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181007 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoX86: Add target description for btver2; make autodetection logic aware of AVX.
Benjamin Kramer [Fri, 3 May 2013 10:20:08 +0000 (10:20 +0000)]
X86: Add target description for btver2; make autodetection logic aware of AVX.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181005 91177308-0d34-0410-b5e6-96231b3b80d8